A dense week. Here is the signal, compressed.
Models
Claude Opus 4.7 ships with 1M context. Anthropic released Opus 4.7 with a 1M token context window at the flagship tier, closing the spec gap with Gemini 1.5 Pro. Pricing stayed flat on a per-token basis, which matters: the marginal cost of long-context calls did not increase. Early reports from engineering teams doing repo-scale reasoning show genuine quality improvements over the previous 200K ceiling. Sub-100K performance is reported as flat versus 4.6, which the Anthropic model card confirms. The practical recommendation: don’t swap your default model, but do revisit your long-document pipelines.
GPT-4o gets an audio latency improvement. OpenAI pushed an update to the real-time API that reduces median audio latency by 30-40ms. This matters for voice application developers where latency perception thresholds are in the 200-300ms range. The improvement narrows the gap with specialized TTS/STT stacks for conversational applications.
Tools
Cursor-Windsurf integration timeline. Following the acquisition close, Cursor published the integration roadmap: Windsurf continues as a standalone product through Q4 2026, with gradual feature migration and a migration tool for existing Windsurf settings and configurations. The user base transition to Cursor is planned for Q1 2027. Teams on annual Windsurf contracts are being honored at renewal.
LangGraph 0.3 release. The stateful agent framework from LangChain released version 0.3 with improved checkpoint persistence, better support for parallel node execution, and a revised streaming API. If you’re building multi-step agents and haven’t evaluated LangGraph since 0.1, the current version is worth looking at — the 0.1 and 0.2 releases had persistent state bugs that made production deployment unreliable.
Applied
The enterprise RAG conversation is getting more honest. A pattern in practitioner discussions this week: more teams are publicly acknowledging that their RAG pilots failed — and being specific about why. The dominant failure mode being cited is knowledge base quality, not retrieval architecture. This is consistent with what we’ve reported on the 70% pilot failure rate. The shift to honest post-mortems is useful for the industry; it moves the conversation from “RAG works/doesn’t work” to “here are the specific conditions under which it works.”
GitHub Copilot Business hits 50K enterprise customers. Microsoft’s Q2 earnings call included a figure: GitHub Copilot Business now has over 50,000 enterprise customers. At $19/seat/month, the revenue run rate is substantial. The number validates the coding-assistant category at enterprise scale, even as startups compete on quality for specific use cases.
Policy
EU Act compliance window: 4 months remain. The 12-month compliance window for general-purpose AI model obligations ends in August 2026. Four months is not much time for organizations that haven’t started. The training data summary requirement is the binding constraint for most US providers — the Commission’s template requires source-level breakdown, not aggregate categories. If you’re building on top of a frontier model and you have EU revenue, you need to understand your provider’s compliance posture.
Colorado AI bill passes second chamber. Colorado’s AI accountability bill passed its second chamber with amendments narrowing the high-risk category definitions. It goes to the governor’s desk next week. The Colorado bill is the closest thing to EU-style AI regulation in a US state — if it’s signed and survives legal challenge, it sets a template that other state legislatures have been watching.
That’s the week. The long-form piece on the inference cost curve published Wednesday. The open vs. closed model framing piece published Friday. Both are in the analysis section.